Professional Documents
Culture Documents
Textbook A Survey of Fractal Dimensions of Networks Eric Rosenberg Ebook All Chapter PDF
Textbook A Survey of Fractal Dimensions of Networks Eric Rosenberg Ebook All Chapter PDF
https://textbookfull.com/product/fractal-functions-dimensions-
and-signal-analysis-santo-banerjee/
https://textbookfull.com/product/a-survey-of-modern-english-
stephan-gramley/
https://textbookfull.com/product/a-survey-of-mathematics-10th-
edition-allen-r-angel/
https://textbookfull.com/product/the-fractal-dimension-of-
architecture-michael-j-ostwald/
A Brief Survey of Quantitative EEG 1st Edition Kaushik
Majumdar
https://textbookfull.com/product/a-brief-survey-of-quantitative-
eeg-1st-edition-kaushik-majumdar/
https://textbookfull.com/product/survey-of-operating-systems-
jane-holcombe/
https://textbookfull.com/product/survey-of-operating-systems-
jane-holcombe-2/
https://textbookfull.com/product/fractal-functions-fractal-
surfaces-and-wavelets-second-edition-peter-r-massopust/
https://textbookfull.com/product/technology-critical-history-of-
a-concept-eric-schatzberg/
SPRINGER BRIEFS IN COMPUTER SCIENCE
Eric Rosenberg
A Survey
of Fractal
Dimensions of
Networks
123
SpringerBriefs in Computer Science
Series editors
Stan Zdonik, Brown University, Providence, Rhode Island, USA
Shashi Shekhar, University of Minnesota, Minneapolis, Minnesota, USA
Xindong Wu, University of Vermont, Burlington, Vermont, USA
Lakhmi C. Jain, University of South Australia, Adelaide, South Australia, Australia
David Padua, University of Illinois Urbana-Champaign, Urbana, Illinois, USA
Xuemin Sherman Shen, University of Waterloo, Waterloo, Ontario, Canada
Borko Furht, Florida Atlantic University, Boca Raton, Florida, USA
V. S. Subrahmanian, University of Maryland, College Park, Maryland, USA
Martial Hebert, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA
Katsushi Ikeuchi, University of Tokyo, Tokyo, Japan
Bruno Siciliano, Università di Napoli Federico II, Napoli, Italy
Sushil Jajodia, George Mason University, Fairfax, Virginia, USA
Newton Lee, Newton Lee Laboratories, LLC, Tujunga, California, USA
More information about this series at http://www.springer.com/series/10028
Eric Rosenberg
A Survey of Fractal
Dimensions of Networks
123
Eric Rosenberg
AT&T Labs
Middletown, NJ, USA
© The Author(s), under exclusive licence to Springer International Publishing AG, part of Springer
Nature 2018
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer International Publishing AG part
of Springer Nature.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To Solomon and Asher
“as an ook cometh of a litel spyr”
from “Troilus and Criseyde” (1374)
by Geoffrey Chaucer
as read by (if you listen very closely)
your great-grandfather
Preface
vii
viii Preface
Many thanks to Kartik Pandit and Curtis Provost, and a special acknowledgment
to Robert Murray, for their comments and suggestions on this survey. Many thanks
also to Paul Drougas, Senior Editor at Springer, and to Frank Politano, Esq of K&L
Gates LLP, for bringing this brief to fruition.
ix
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Tables of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Box Counting and Hausdorff Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Covering a Complex Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Box Counting with Diameter-Based or Radius-Based Boxes . . . . . . . . 9
3 Network Box Counting Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1 Node Coloring Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Node Coloring for Weighted Networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Random Sequential Node Burning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Set Covering Formulation and a Greedy Method . . . . . . . . . . . . . . . . . . . . . 19
3.5 Box Burning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6 Box Counting for Scale-Free Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4 Lower Bounds on Box Counting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1 Mathematical Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Dual Ascent and Dual Adjustment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.3 Bounding the Fractal Dimension. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5 Correlation Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6 Mass Dimension for Infinite Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.1 Transfinite Fractal Dimension. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7 Volume and Surface Dimensions for Infinite Networks . . . . . . . . . . . . . . . . . 51
8 Information Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
9 Generalized Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
10 Non-monotonicity of Generalized Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
11 Zeta Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
xi
Chapter 1
Introduction
Consider the network G = (N, A) where N is the set of nodes connected by the set
A of arcs. Let N ≡ |N| be the number of nodes, and let A ≡ |A| be the number
of arcs. (We use “≡” to denote a definition.) Unless otherwise specified, henceforth
we assume that G is a “complex network”: an arbitrary network without special
structure (as opposed to, e.g., a regular lattice), for which all arcs have unit cost (so
the length of a shortest path between two nodes is the number of arcs in that path),
and all arcs are undirected (so the arc between nodes i and j can be traversed in
either direction). We assume that G is connected, meaning there is a path of arcs
in A connecting any two nodes in N. Many different measures have been used to
describe a complex network [7]. For example,
1. the density of G is A/[N(N −1)/2], which is the ratio of the number of arcs in G
to the number of arcs that would be present if each pair of nodes were connected
by an arc;
2. the diameter Δ of G is defined by Δ ≡ max{ dist(x, y) | x, y ∈ N }, where
dist(x, y) is the length of the shortest
path between nodes x and y;
3. the average node degree is (1/N) n∈N δn , where δn is the node degree of node
n (the number of arcs incident to node n).
Another way to describe G is to compute a fractal dimension of G. The interest
in computing fractal dimensions of networks began about 2003 [10] (although
one important paper [38] appeared as early as 1988) and was inspired by the
study of fractal dimensions of geometric objects. For geometric objects, the fractal
dimensions of interest usually have non-integer values, and for that reason a fractal
is sometimes defined as something with a non-integer dimension. However, a good
definition of “a fractal” has been an elusive goal, and research has instead focused
on the theory and application of different fractal dimensions. Similarly, we are not
concerned here with a definition of “a fractal network”. Rather, our goal is to survey
the most important of the different fractal dimensions that have been proposed for
networks.
This book is organized as follows. In this chapter we will present the box
counting and Hausdorff dimensions. Chapter 2 considers the covering of a complex
network by boxes. Several heuristics for box counting are presented in Chap. 3, and
Chap. 4 is concerned with lower bounds on box counting. The correlation dimension
is presented in Chap. 5. The mass dimension and the transfinite fractal dimension of
an infinite network are studied in Chap. 6, and the volume and surface dimensions of
an infinite network are studied in Chap. 7. The information dimension is the subject
of Chap. 8. Generalized dimensions and the sandbox method are studied in Chap. 9.
The non-monotonicity of the generalized dimensions is explored in Chap. 10. The
final chapter, Chap. 11, presents the zeta dimension.
Many of the fractal dimensions studied in this book are defined in terms of a
covering of G by “boxes”, and the reason for covering G by boxes is that this is
how the fractal dimension of a geometric object was first defined. So we start, in
the next section, with a review of the box counting and Hausdorff dimensions of a
geometric object. For convenience, some symbols frequently used in this survey are
summarized in Table 1.1, and a list of the fractal dimensions we study is provided in
Table 1.2. The end of a definition, or proof, or example, is indicated by the symbol
.
The simplest fractal dimension is the box counting dimension, which is based on
covering a geometric object Ω ⊂ RE by equal sized E-dimensional hypercubes.
For example, for E = 1, consider a line segment of length L. If we measure the line
segment using a ruler of length s, where s L, the number B(s) of rule lengths
needed is given by B(s) ≈ Ls −1 . We call B(s) the “number of boxes” of size s
needed to cover the segment. For E = 1, a “box” of size s is a line segment of
length s. Since the exponent of s in Ls −1 is −1, we say that a line segment has
a box counting dimension of 1. Now consider a two-dimensional square with side
length L. If we cover the square by small squares of side length s, where s L, the
number B(s) of small squares needed is given by B(s) ≈ L2 s −2 . Since the exponent
of s in L2 s −2 is −2, we say that the square has box counting dimension 2.
To provide a general definition of the box counting dimension of a geometric
object, let Ω be a closed and bounded subset of RE . By the “linear size” of Ω
we mean the diameter of Ω (the maximal Euclidean distance between any two
points of Ω, denoted by diam(Ω)), or the maximal variation in any coordinate (i.e.,
maxx,y∈Ω max1≤i≤E |xi − yi |). Let s 1 be the linear size of a small box, where a
“box” is an E-dimensional hypercube. By a box of size s we mean a box of linear
size s. A set of boxes covers Ω if each point in Ω belongs to at least one box.
1.2 Box Counting and Hausdorff Dimensions 3
Definition 1.1 Let B(s) be the minimal number of boxes of size s needed to
cover Ω. If
log B(s)
lim (1.1)
s→0 log(1/s)
exists, then the limit is called the box counting dimension of Ω and is denoted
by dB .
Roughly speaking, if dB is the box counting dimension of Ω then B(s) behaves
as s −dB for s 1. In practice, the computation of dB typically begins by selecting
a set {s1 , s2 , · · · , sK } of box sizes. For each value of s, we determine the minimal
number B(s) of boxes of size s needed to cover Ω. By plotting log B(s) vs. log s
for the K values of s, a range of s can be identified over which the plot is roughly
linear [3, 27, 28, 45]. Then dB can be determined, e.g., by linear regression.
Although the limit (1.1) may not exist [12], the lim inf and lim sup always exist.
The lower box counting dimension dB is defined by
log B(s)
dB ≡ lim inf , (1.2)
s→0 log (1/s)
log B(s)
dB ≡ lim sup . (1.3)
s→0 log (1/s)
J
d
v(d, s) ≡ inf diam(Xj ) , (1.4)
C (s)
j =1
where the infimum is over all s-coverings C (s) of Ω. We take the infimum since the
goal is to cover Ω with small sets Xj as efficiently as possible.
We can think of v(d, s) as the d-dimensional volume of Ω. For almost all values
of d, the limit lims→0 v(d, s) is either 0 or ∞, where by ∞ we mean +∞. For
example, suppose we cover the unit square [0, 1] × [0, 1] by small squares of side
1.2 Box Counting and Hausdorff Dimensions 5
√
length s. We need 1/s 2 small squares, the diameter of each square is 2s, and
√ √ d
v(d, s) = (1/s 2 )( 2s)d = 2 s d−2 . We have
⎧
⎪
⎨∞
⎪ if d < 2
lim s d−2 = 1 if d = 2
s→0 ⎪
⎪
⎩0 if d > 2.
Thus, for example, if d = 3 then the unit square [0, 1] × [0, 1] has zero volume; if
d = 1 then the unit square has infinite length.
For a given d, as s decreases, the set of available covers shrinks, so v(d, s)
increases as s decreases. Thus
always exists in [0, ∞) ∪ {∞}; that is, v (d) might be ∞. (We call v (d) the d-
dimensional Hausdorff measure of Ω.) Since for each fixed s < 1 the function
v(d, s) is non-increasing with d, then v (d) is also non-increasing with d. For d ≥ 0
and d ≥ 0, definition (1.5) implies [53]
These two assertions imply the existence of a unique value of d, called the Hausdorff
dimension of Ω and denoted by dH , such that v (d) = ∞ for d < dH and v (d) = 0
for d > dH . Formally,
The previous chapter showed how the box counting and Hausdorff dimensions of a
geometric object Ω are computed from a covering of Ω. With this background, we
can now consider what it means to cover a complex network G, and how a fractal
dimension can be computed from a covering of G. We require some definitions. The
network B is a subnetwork of G if B can be obtained from G by deleting nodes and
arcs. By a box we mean a subnetwork of G. A box is disconnected if some nodes in
the box cannot be connected by arcs in the box. Let {Bj }Jj=1 ≡ {B1 , B2 , · · · , BJ }
be a collection of boxes. Two types of coverings of G have been proposed: node
coverings and arc coverings. Let s be a positive integer.
Definition 2.1 (i) The set {Bj }Jj=1 is a node s-covering of G if for each j we have
diam(Bj ) < s and if each node in N is contained in exactly one Bj . (ii) The set
{Bj }Jj=1 is an arc s-covering of G if for each j we have diam(Bj ) < s and if each
arc in A is contained in exactly one Bj .
If Bj is a box in a node or arc s-covering of G then the requirement diam(Bj ) < s
in Definition 2.1 implies that Bj is connected. However, this requirement, which is
a standard assumption in defining the box counting dimension of G (e.g., [16, 29,
30, 48, 56]), may frequently be violated, for good reasons, in some methods for
determining the fractal dimensions of G, as we will discuss in Sect. 3.6.
It is possible to define a node covering of G to allow a node to be contained
in more than one box; coverings with possibly overlapping boxes are studied in
[15, 60]. The great advantage of non-overlapping boxes is that they immediately
yield a probability distribution, as discussed in Chap. 8. The probability distribution
obtained from a non-overlapping node covering of G is the basis for computing
the information dimension dI and the generalized dimensions Dq of G (Chap. 9).
Therefore, in this survey, each node covering of G is assumed to use non-
overlapping boxes, as specified in Definition 2.1.
There are two main approaches used to define boxes for use in covering G: diameter-
based boxes and radius-based boxes.
Definition 2.3 (i) A radius-based box G(n, r) with center node n ∈ N and radius r
is the subnetwork of G containing all nodes whose distance to n does not exceed r.
Let BR (r) be the minimal number of radius-based boxes of radius at most r needed
to cover G. (ii) A diameter-based box G(s) of size s is a subnetwork of G of diameter
s − 1. Let BD(s) denote the minimal number of diameter-based boxes of size at most
s needed to cover G.
Thus the node set of G(n, r) is {x ∈ N | dist(n, x) ≤ r}. Radius-based boxes
are used in the Maximum Excluded Mass Burning and Random Sequential Node
Burning methods described in Chap. 3. Interestingly, the above definition of a
radius-based box may frequently be violated in the Maximum Excluded Mass
Burning and Random Sequential Node Burning methods. In particular, some radius-
based boxes created by those methods may be disconnected, or some boxes may
contain only some of the nodes whose distance to the center node n does not
exceed r.
A diameter-based box G(s) is not defined in terms of a center node; instead,
for x, y ∈ G(s) we require dist(x, y) < s. Diameter-based boxes are used in
the Box Burning and Compact Box Burning heuristics described in Chap. 3. The
above definition of a diameter-based box also may frequently be violated in the
Box Burning and Compact Box Burning methods. Also, since each node in G must
belong to exactly one Bj in an s-covering {Bj }Jj=1 using diameter-based boxes, then
in general we will not have diam(Bj ) = s − 1 for all j . To see this, consider a chain
of three nodes (call them x, y, and z), and let s = 2. The minimal 2-covering using
diameter-based boxes requires two boxes, B1 and B2 . If B1 covers x and y then B2
covers only z, so the diameter of B2 is 0.
The minimal number of diameter-based boxes of size at most 2r + 1 needed to
cover G is, by definition, BD(2r + 1). We have BD(2r + 1) ≤ BR (r) [29]. To see
this, let G(nj , rj ), j = 1, 2, · · · , BR (r) be the boxes in a minimal covering of G
using radius-based boxes of radius at most r. Then rj ≤ r for all j . Pick any j , and
consider box G(nj , rj ). For any nodes x and y in G(nj , rj ) we have
so G(nj , rj ) has diameter at most 2r. Thus these BR (r) boxes also serve as a covering
of size 2r + 1 using diameter-based boxes. Therefore, the minimal number of
diameter-based boxes of size at most 2r +1 needed to cover G cannot exceed BR (r);
that is, BD(2r + 1) ≤ BR (r).
10 2 Covering a Complex Network
The reverse inequality does not in general hold, since a diameter-based box of
size 2r +1 can contain more nodes than a radius-based box of radius r. For example,
consider the network G of Fig. 2.3. The only nodes adjacent to n are x and z, so
G(n, 1) = {n, x, z} and BR (1) = 2. Yet the diameter of G is 2, so it can be covered
by a single diameter-based box of size 3, namely G itself, so BD(3) = 1. Thus BR (r)
and BD(2r + 1) are not in general equal. Nonetheless, for the C. elegans and Internet
backbone networks studied in [56], the calculated fractal dimension was the same
whether radius-based or diameter-based boxes were used. Similarly, both radius-
based and diameter-based boxes yielded a fractal dimension of approximately 4.1
for the WWW (the World Wide Web) [29].
The term box counting refers to computing a minimal s-covering of G for a
range of values of s, using either radius-based boxes or diameter-based boxes.
Conceivably, other types of boxes might be used to cover G. In the fractal literature,
the box counting dimension dB is often informally defined by the scaling BD(s) ∼
s −dB . (The symbol “∼”, frequently used in the fractal literature but often with
different meanings, should here be interpreted to mean “approximately behaves
like”.) Definition 2.4 below provides a more computationally useful definition of
dB for a complex network.
Definition 2.4 G has box counting dimension dB if over some range of s and for
some constant c we have
Alternatively, (2.1) can be written as log BD(s) ≈ −dB log s + c. If G has box
counting dimension dB then over some range of s we have BD(s) ≈ as −dB for some
constant a. In the terminology of [16], if the box counting dimension for G exists,
then G enjoys the fractal scaling property, or, more simply, G is fractal. The main
feature apparently displayed by fractal networks is a repulsion between hubs, where
a hub is a node with a significantly higher node degree than a non-hub node. That
is, the highly connected nodes tend to be not directly connected [72]. This tendency
can be quantified using the joint node degree distribution p(δ1 , δ2 ) that a node with
degree δ1 and a node with degree δ2 are neighbors (i.e., connected by a single arc).
In contrast, for a non-fractal network G, hubs are mostly connected to other hubs,
which implies that G enjoys the small-world property [16]. (Roughly speaking, G
is a small-world network if diam(G) grows as log(N ) [57].) Also, the concepts of
2.1 Box Counting with Diameter-Based or Radius-Based Boxes 11
modularity and fractality for a network are closely related. Interconnections within
a module (e.g., a biological sub-system) are more prevalent than interconnections
between modules. Similarly, in a fractal network, interconnections between a hub
and non-hub nodes are more prevalent than interconnections between hubs. Non-
fractal networks are typically characterized by a sharp decay of BD(s) with s, which
is better described by an exponential law BD(s) ∼ e−β s , where β > 0, rather than
by a power law BD(s) ∼ s −β , with a similar statement holding if radius-based boxes
are used. These two cases are illustrated in Fig. 2.4, taken from [16], where the solid
circles are measurements from a fractal network, and the hollow circles are from a
non-fractal network.
Chapter 3
Network Box Counting Heuristics
diameter-based box of size s. It follows that BD(s) ≤ χ (Gs ). To prove the reverse
inequality, consider any minimal s-covering using BD(s) boxes, and let B be any
box in this covering. For any nodes x and y in this box we have dist(x, y) < s, so
x and y are not connected by an arc in Gs . Thus x and y can be assigned the same
color, which implies χ (Gs ) ≤ BD(s).
We illustrate the node coloring formulation using the network of Fig. 3.1. For s =
3, the auxiliary graph G3 is given by Fig. 3.2a. Arc (x, y) exists in G3 if dist(x, y) ≥
3. Thus node c is isolated in G3 since its distance in G to all other nodes does not
exceed 2. Also, the distance in G from g to all nodes except a does not exceed 2, so
arc (g, a) exists in G3 .
The chromatic number χ (G3 ) of the simple network of Fig. 3.2a can be exactly
computed using the Greedy Coloring method [56], which assigns colors based on a
random ordering of the nodes. Typically Greedy Coloring would be run many times,
using different random orderings of the nodes; using 10,000 random orderings,
Greedy Coloring has been shown to provide significant accuracy. Moreover the
method is very efficient, since, for a given random ordering of the nodes, a single
pass through all the nodes suffices to compute an s-covering of G for all box sizes
s [56].
We illustrate Greedy Coloring using Fig. 3.2b. Suppose we randomly pick a as
the first node, and assign the color yellow to node a (a yellow node is indicated
using a small square box). Then d, e, f , and g cannot be colored yellow, so we
color them blue (a blue node is indicated using a small oval box). We can color
b yellow since it is connected only to nodes already colored blue. Since c is
isolated we are free to assign it any color, so we color it yellow. We are done;
nodes a, b, and c are in the yellow box and nodes d, e, f , and g are in the blue
box. This is an optimal coloring, since at least two colors are needed to color any
graph with at least one arc. Figure 3.3 illustrates, in the original network, the two
boxes in this minimal covering for s = 3.
For s = 4, the auxiliary graph G4 is shown in Fig. 3.4a. There is an arc (x, y)
in G4 if in G we have dist(x, y) ≥ 4. We again apply the Greedy Coloring
heuristic to compute the chromatic number χ (G4 ). Suppose we randomly pick a
as the first node, and assign the color yellow to node a. Then e and f cannot
be colored yellow, so we color them blue. The remaining nodes are isolated
so we arbitrarily color them blue. We are done; a is in the yellow box and the
remaining nodes are in the blue box. This is also an optimal coloring. Figure 3.4b
illustrates, in the original network, the two boxes in the minimal covering for s = 4.
3.2 Node Coloring for Weighted Networks 15
Fig. 3.2 (a) Auxiliary graph with s=3, and (b) its coloring
Fig. 3.4 (a) Auxiliary graph with s=4, and (b) its coloring
For an unweighted network, the distance between two nodes (also known as the
chemical distance, or the hop count) ranges from 1 to the diameter Δ of the network.
However, when applying box counting to a weighted network, choosing box sizes
between 2 and Δ will not in general be useful. For example, if the network diameter
is less than 1, then the entire network is contained in a box of size 1. One simple
approach to dealing with box size selection for weighted networks is to multiply
each arc length by a sufficiently large constant k. For example, if we approximate
each arc length by a rational number, then choosing k to be the least common
denominator of all these rational numbers will yield a set of integer arc lengths.
16 3 Network Box Counting Heuristics
Even if all the arc lengths are integer, a set of box sizes must still be selected. The
box sizes could be selected using a default method, such as starting with a box
size less than the network diameter, and decreasing the box size by a factor of 2
in each iteration. Alternatively, the box sizes could be determined by an analysis
of the set of arc lengths. This is the approach taken in [69], and we describe their
method using the network of Fig. 3.5. The nodes are a, b, c, d, e, f , and each arc
length is shown. We first pre-process the data by computing the shortest distance
dij between each pair (i, j ) of nodes. The largest dij is the diameter Δ. For this
example we have Δ = 1.16, which is the distance between nodes a and c. The
second pre-processing step is to sort all the dij values in increasing order. For this
example, the five smallest dij values are 0.05, 0.1, 0.3, 0.4, 0.5. Next we compute
the successive sums of the ordered dij values, stopping when the sum first exceeds
Δ. The first sum is σ (1) = 0.05, the second is σ (2) = σ (1) + 0.1 = 0.15, the third
is σ (3) = σ (2) + .3 = .45, the fourth is σ (4) = σ (3) + 0.4 = 0.85, and finally
σ (5) = σ (4) + 0.5 = 1.35 > Δ. We set s1 = σ (4) = 0.85 since this is the largest
sum not exceeding Δ.
Next we create an auxiliary graph G such that an arc in G exists between nodes
i and j if dij ≥ s1 . There are four pairs of nodes for which dij ≥ s1 , namely (a, c),
(a, e), (a, f ), and (c, b), so G has four arcs. Node d does not appear in G. The
length of arc (i, j ) in G is dij , e.g., daf = 0.95, which is the length of the shortest
path in G from a to f . Next we assign a weight to each node in G. For node i in G,
the weight w(i) is
w(i) ≡ dij .
(i,j )∈G
The w(i) values are the underlined values in Fig. 3.6 next to each node. Thus w(a) =
0.9 + 1.16 + 0.95 = 3.01 for the three arcs in G incident to a, and w(f ) = 0.95 for
the one arc in G incident to f . As in Sect. 3.1, we color the nodes of G so that the
endpoints of each arc in G are assigned different colors. Each color will correspond
to a distinct box, so using the minimal number of colors means using the fewest
boxes. We start with the node with the highest weight. This is node a, whose weight
is 3.01. Suppose we assign to a the color yellow, as indicated by the node name
in a small square, as shown in Fig. 3.7. Then nodes c, e, and f cannot be colored
yellow, so we color them blue, as indicated by the node name in a small circle.
The remaining node in G to be colored is node b, and it can be colored yellow.
3.3 Random Sequential Node Burning 17
The final step in this iteration is to color to each node not in G. If node i does not
appear in G then, by construction of G, we have dij < s1 for j ∈ N. Thus i can be
assigned to any nonempty box, so i can be assigned any previously used color. In
our example, only d is not in G, and we arbitrarily color it yellow. Thus for the
initial box size s1 we require only two boxes, a yellow box containing a, b, and d,
and a blue box containing c, e, and f . This concludes the iteration for the initial
box size s1 .
For the second iteration, we first need the new box size s2 . The method of [69]
simply takes the next smallest of the sums. Since s1 = σ (4), then s2 = σ (3) = 0.45.
With this new box size, we continue as before, creating a new auxiliary graph G
containing each arc (i, j ) such that dij ≥ s2 , determining the weight of each node
in G, coloring the nodes of G, and then coloring the nodes not in G. Then we select
s3 = σ (2) and continue in this manner.
This above material is one of the few sections in this survey concerning weighted
networks. We now return to the study of unweighted networks, and examine other
methods for box counting.
In this section we study the Random Sequential Node Burning method of [29] for
covering G using radius-based boxes. For a given radius r, the procedure is as
follows. Initially all nodes are uncovered (or “unburned”, in the terminology of [29],
i.e., not yet assigned to a box), and the box count BR (r) is initialized to 0. In each
iteration, we first pick a random node n which may be covered or uncovered, but
which has not previously been selected as the center node of a box. We create the
new radius-based box G(n, r). Next we add to G(n, r) each uncovered node whose
distance from n does not exceed r. If G(n, r) contains no uncovered nodes, then
18 3 Network Box Counting Heuristics
1 if dist(i, j ) ≤ r,
Mijr = (3.1)
0 otherwise .
(The superscript r does not mean the r-th power of the matrix M.) For an undirected
graph, M r is symmetric. For example, for r = 1, the matrix M 1 corresponding to the
network of Fig. 3.9 is the same as the node-node incidence matrix of the network,
namely
⎛ ⎞
1100001
⎜1 1 1 0 0 0 0⎟
⎜ ⎟
⎜0 1 1 1 0 0 1⎟
⎜ ⎟
⎜ ⎟
M1 = ⎜ 0 0 1 1 1 0 0 ⎟ .
⎜ ⎟
⎜0 0 0 1 1 1 0⎟
⎜ ⎟
⎝0 0 0 0 1 1 1⎠
1010011
For the same network and r = 2 we have
⎛ ⎞
111 0 0 1 1
⎜1 1 1 1⎟
⎜ 1 0 0 ⎟
⎜1 1 1 1⎟
⎜ 1 1 1 ⎟
⎜ ⎟
M = ⎜0 1 1
2
1 1 1 1⎟ .
⎜ ⎟
⎜0 0 1 1 1 1 1⎟
⎜ ⎟
⎝1 0 1 1 1 1 1⎠
111 1 1 1 1
20 3 Network Box Counting Heuristics
The minimal number BR (r) of boxes needed to cover G is the optimal objective
function value of the following binary integer program (an integer program is a
linear optimization problem whose variables are restricted to be integer valued):
N
minimize xn (3.2)
n=1
N
subject to Mjr n xn ≥ 1 for j = 1, 2, · · · , N (3.3)
n=1
xn = 0 or 1 for n = 1, 2, · · · , N. (3.4)
Here Mjr n is the element in row j and column n of the matrix M r . The objective
function (3.2) is the number of center nodes used in the covering, i.e., the number
of boxes used in the covering. The left hand side of constraint (3.3) is the number of
boxes covering node j , so this constraint requires that each node be within distance
r of at least one center node used in the covering.
To express this formulation more compactly, let x = (x1 , x2 , · · · , xN ) be the
column vector of size N (that is, a matrix with N rows and one column) and let
x T be the transpose of x (so x T is a matrix with one row and N columns). Let 1
= (1, 1, · · · , 1) be the column vector of size N each of whose components is 1.
Then the above set covering formulation can be written as
minimize x T 1
subject to M r x ≥ 1
xn = 0 or 1 for n = 1, 2, · · · , N.
Another random document with
no related content on Scribd:
innostukseni, kaikkeni! Mutta minulla on rohkeutta, minulla on
voimaa raivata itselleni uusi tie. Kunpa Jumala auttaisi minua, ja
Hän sen kyllä tekee. Se on kohtaloni, tunnen sen, minun pitää
seurata sitä, minun pitää taipua sen alle, minulla ei ole voimaa
vastustaa sitä. Itken ajatellessani, mitä minä jätän. Riemuitsen
ajatellessani itseni vapaaksi. Nyt kai aavistatte, mitä minä teen.
Toivon, ettette ilmoita sitä muille. Älkää halveksiko minua. Teen
sen, koska minun luonnollani on mahdotonta tehdä toisin!
Iida Aalberg.»
*****
Ida Aalberg sai, vaikka erosikin Suomalaisesta teatterista, keväällä
1883 pienen valtioavun ulkomaamatkaa varten. Hän suunnitteli
esiintymistä Kööpenhaminassa, jonne Z. Topelius häntä lämpimästi
suosittelu Sarah Bernhardt vieraili näihin aikoihin Tukholmassa ja
Kööpenhaminassa, ja uskotulleen Bertha Forsmanille Ida Aalberg
kertoo ensi vaikutelmistaan seuraavalla tavalla:
Teidän
Ida Aalberg.»
Ida Aalberg istui hyvin monena iltana Porte Saint Martin teatterissa
ja seurasi suuren Sarah'n näyttelemistä. Ei ole miksikään häpeäksi
hänelle, että hän otti vastaan vaikutteita itseään vanhemmalta
suurelta taiteilijalta. Saihan Eleonora Dusekin kiittää Sarah’ta tiensä
viitoittamisesta.
Kotimaahan Ida Aalberg lähetti ärtyneitä kirjeitä kuullessaan
juorupuheista, joita hänestä siellä levitettiin. »Jumaloin Suomeani,
mutta inhoan ihmisiä siellä», hän kirjoittaa. Huhut koskivat hänen
luultua jäämistään Ranskaan, Sarah Bernhardt muka oli tarjonnut
hänelle paikan teatterissaan. Toiset tiesivät hänen aikovan mennä
avioliittoon ja senvuoksi luopuvan Suomalaisesta teatterista. Vanha
kaulatauti vaivasi häntä myöskin Seinen kaupungissa ja teki hänet
toisinaan melkein epätoivoiseksi:
*****
*****
NEUVONANTAJIA JA YSTÄVIÄ.